Optimizing Training Population Size and Content to Improve Prediction Accuracy of FHB-Related Traits in Wheat

Abstract

Genomic selection combines phenotypic and molecular marker data from a training population to predict the genotypic values of untested lines. It can improve breeding efficiency as large pools of untested lines can be evaluated for selection. Training population (TP) composition is one of the most important factors affecting the accuracy of genomic prediction. The University of Minnesota wheat breeding program implements genomic selection at the F5 stage for Fusarium head blight (FHB) resistance. This study used field data for FHB resistance in wheat (Triticum aestivum L.) to investigate the use of small-size TPs designed with and without stratified sampling for three FHB traits in three different F5 populations (TP17, TP18, and TP19). We also compared the accuracies of these two TP design methods with the accuracy obtained from a large size TP. Lastly, we evaluated the impact on trait predictions when the parents of F5 lines were included in the TP. We found that the small size TP selected randomly, without stratification, had the lowest predictive ability across the three F5 populations and across the three traits. This trend was statistically significant (p = 0.05) for all three traits in TP17 and two traits in TP18. Designing a small-size TP by stratified sampling led to a higher accuracy than a large-size TP in most traits across TP18 and TP19; this is because stratified sampling allowed the selection of a small set of closely related lines. We also observed that the addition of parental lines to the TP and evaluating the TP in two replications led to an increase in predictive abilities in most cases.

Training Set Optimization for Sparse Phenotyping in Genomic Selection: A Conceptual Overview

“…Other studies also stressed the importance of considering an other way to construct the TRS by random sampling (Lorenz and Smith, 2015 ; He et al, 2016 ; Cericola et al, 2017 ; Neyhart et al, 2017 ; Norman et al, 2018 ; de Bem Oliveira et al, 2020 ; Olatoye et al, 2020 ), clustering approaches (Akdemir et al, 2015 ; Isidro et al, 2015 ; Bustos-Korts et al, 2016 ; Rincent et al, 2017 ; Norman et al, 2018 ; Guo et al, 2019 ; Sarinelli et al, 2019 ; Adeyemo et al, 2020 ), by using different levels of relatedness between TRS and TS (Lorenz and Smith, 2015 ; Berro et al, 2019 ; Roth et al, 2020 ) or by using other alternatives algorithms to CD-mean and PEV-mean such as different design matrix algorithm (Akdemir and Isidro-Sánchez, 2019 ), estimated theoretical accuracy (EthAcc) (Mangin et al, 2019 ), upper bound reliability (Yu et al, 2020 ), or the Fast and Unique Representative Subset Selection (FURS) (Guo et al, 2014 ). A criterion that is derived directly from Pearson's correlation between GEBVs and phenotypic values of the TS derived from the GBLUP model showed higher predictive ability than CD and PEV (Ou and Liao, 2019 ).…”

Genome‐wide association mapping for field and seedling resistance to the emerging Puccinia graminis f. sp. tritici race TTRTF in wheat

“…Genetic relationship among the lines in the population is discussed as kinship among the lines and was estimated in Tassel 5.2.74 using the Centered_IBS method using default parameters, that is, maximum number of alleles = 6. Presence of population structure and stratification was assessed using principal component (PC) analysis with the function ‘prcomp’ in R. Subpopulations observed during the PC analysis were visualized by grouping the lines using a k ‐means clustering algorithm (Adeyemo et al., 2020).…”

Building a Calibration Set for Genomic Prediction, Characteristics to Be Considered, and Optimization Approaches

“…It was proposed in numerous studies to define the CS by stratified sampling. These algorithms ensure that each group is well represented in the CS, possibly taking the size of each group into account [84,87,[90][91][92][93]. The efficiency of these approaches to increase predictive ability was disappointing, as they were often not better (sometimes worse) than random sampling, and most of the time not as good as relatedness-based criteria not taking structure into account.…”